基于机器学习的灭弧栅片表面缺陷检测方法

Machine Learning-Based Surface Defect Detection Method for Arc Extinguishing Grid

为了实现金属灭弧栅片表面缺陷的自动检测,引入了CCD成像系统并提出了表面缺陷检测三步法:第一步对原始图像进行去噪,第二步将灭弧栅片从背景中提取出来,第三步利用分类器对缺陷产品进行识别。提出了基于方向梯度直方图(HOG)与Gabor特征结合的图像特征提取算法,与传统的基于HOG和基于Gabor特征的算法相比,多分类支持向量机的训练结果显示本方法识别率分别提高了13%和7%。通过设计卷积神经网络框架对缺陷产品进行检测,结果显示正确率为93%。在二分类情况下对支持向量机和卷积神经网络的分类性能进行了比较,结果显示卷积神经网络性能更优。

In order to realize the automatic detection of surface defects of metal arc extinguishing grids,this paper introduced a CCD imaging system and proposed a three-step surface defect detection method:The first step was to denoise the original image,then the second step was to remove the arc grid from the background,the third step used the classifier to identify the defective product.An image feature extraction algorithm based on R-HOG and Gabor features was proposed.Compared with the traditional HOG and Gabor features,the multi-classification support vector machine training results show that the recognition rate of the method proposed in this paper is improved 13%and 7%respectively.The defect products wer e classified by constructing a convolutional neural network framework.The results show that the recognition rate is 93%.The classification performances of SVM and convolutional neural network were compared in the two-category case.The results show that the convolutional neural network has better performance.